Is it actually open source? A layer-by-layer answer

The problem with the phrase

"Open source" gets used about AI as one undifferentiated badge. In practice a working local AI stack has at least four separate layers, each with its own licence, and a product can be genuinely open at two of them and closed at the other two.

So the useful question is not "is it open source?" It is "which parts can I read, which parts can I change, and which parts can I not touch?"

Layer 1: the inference engine

This is the software that loads a model and computes answers. It is the layer where "open source" means the most in day-to-day operation, because it is the thing running on your machine right now.

The engine we present as the open option is MIT-licensed with its source published. You can read it, fork it, build your own, and ship that. That is a real open-source claim with a real licence behind it.

There is also a faster alternative engine available. It is a closed, compiled binary under a custom end-user licence agreement. You may run it commercially and redistribute it unmodified. You may not read its source. We use the phrase "closed source" about it rather than dressing it up.

Layer 2: the model weights

Weights are not source code and they are not free of restrictions either. Each model carries its own licence and they differ substantially.

Some weights are MIT. Some are Apache-2.0. Both are permissive: use them commercially, modify them, redistribute them.

One of the fastest models we support is under the Qwen Community License 1.0, and that one has a clause worth understanding. It restricts serving the model as a hosted service to third parties above a scale threshold. Running it on your own box, for your own work, is fine. Building a business where you run inference on behalf of other people using that model is where you need to read the licence carefully.

We ship the box with the customer running everything locally. That sits on the safe side of the line, and it is a deliberate design choice rather than an accident of how the licensing worked out.

Layer 3: the operating system

Fedora Linux. Free and open source throughout. Everything in it is inspectable and rebuildable. This is the least controversial layer and the most thoroughly established.

Layer 4: the hardware

The chassis CAD and module documentation are open. The mainboard is only partly documented, and the processor is soldered down, so you cannot swap it.

That is "partial" and it is worth saying plainly. You can print your own replacement panels. You cannot fabricate your own motherboard from the published material. Anyone describing repairable-and-open hardware should be specific about which parts, because "open hardware" as a blanket phrase oversells what a soldered APU means for long-term repairability.

How to read a vendor's open-source claim

Ask three questions and the marketing falls away quickly:

1. Which layer? Engine, weights, OS, or hardware. A claim that names none of them is not a claim you can check.
2. Which licence, by name? MIT, Apache-2.0, a community licence, or "custom." "Open" is not a licence.
3. Can you point me at the source? A real open-source claim has a repository with a commit history. If the answer is a marketing page, it is not open.

We publish the full per-layer licence table on the site rather than asking you to take the word for it. If any row in that table changes, we change the row.

Why bother being this specific

Because the vague version of the claim is fragile. A customer who discovers that "open source AI" meant open at one layer and closed at three others stops trusting everything else you said — including the parts that were true.

Precision costs nothing here and buys the only thing that is hard to get back: being believed.